2 * kernel/stop_machine.c
4 * Copyright (C) 2008, 2005 IBM Corporation.
5 * Copyright (C) 2008, 2005 Rusty Russell rusty@rustcorp.com.au
6 * Copyright (C) 2010 SUSE Linux Products GmbH
7 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
9 * This file is released under the GPLv2 and any later version.
11 #include <linux/completion.h>
12 #include <linux/cpu.h>
13 #include <linux/init.h>
14 #include <linux/kthread.h>
15 #include <linux/export.h>
16 #include <linux/percpu.h>
17 #include <linux/sched.h>
18 #include <linux/stop_machine.h>
19 #include <linux/interrupt.h>
20 #include <linux/kallsyms.h>
21 #include <linux/smpboot.h>
22 #include <linux/atomic.h>
25 * Structure to determine completion condition and record errors. May
26 * be shared by works on different cpus.
28 struct cpu_stop_done {
29 atomic_t nr_todo; /* nr left to execute */
30 bool executed; /* actually executed? */
31 int ret; /* collected return value */
32 struct completion completion; /* fired if nr_todo reaches 0 */
35 /* the actual stopper, one per every possible cpu, enabled on online cpus */
38 bool enabled; /* is this stopper enabled? */
39 struct list_head works; /* list of pending works */
42 static DEFINE_PER_CPU(struct cpu_stopper, cpu_stopper);
43 static DEFINE_PER_CPU(struct task_struct *, cpu_stopper_task);
44 static bool stop_machine_initialized = false;
46 static void cpu_stop_init_done(struct cpu_stop_done *done, unsigned int nr_todo)
48 memset(done, 0, sizeof(*done));
49 atomic_set(&done->nr_todo, nr_todo);
50 init_completion(&done->completion);
53 /* signal completion unless @done is NULL */
54 static void cpu_stop_signal_done(struct cpu_stop_done *done, bool executed)
58 done->executed = true;
59 if (atomic_dec_and_test(&done->nr_todo))
60 complete(&done->completion);
64 /* queue @work to @stopper. if offline, @work is completed immediately */
65 static void cpu_stop_queue_work(unsigned int cpu, struct cpu_stop_work *work)
67 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
68 struct task_struct *p = per_cpu(cpu_stopper_task, cpu);
72 spin_lock_irqsave(&stopper->lock, flags);
74 if (stopper->enabled) {
75 list_add_tail(&work->list, &stopper->works);
78 cpu_stop_signal_done(work->done, false);
80 spin_unlock_irqrestore(&stopper->lock, flags);
84 * stop_one_cpu - stop a cpu
86 * @fn: function to execute
87 * @arg: argument to @fn
89 * Execute @fn(@arg) on @cpu. @fn is run in a process context with
90 * the highest priority preempting any task on the cpu and
91 * monopolizing it. This function returns after the execution is
94 * This function doesn't guarantee @cpu stays online till @fn
95 * completes. If @cpu goes down in the middle, execution may happen
96 * partially or fully on different cpus. @fn should either be ready
97 * for that or the caller should ensure that @cpu stays online until
98 * this function completes.
104 * -ENOENT if @fn(@arg) was not executed because @cpu was offline;
105 * otherwise, the return value of @fn.
107 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
109 struct cpu_stop_done done;
110 struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
112 cpu_stop_init_done(&done, 1);
113 cpu_stop_queue_work(cpu, &work);
114 wait_for_completion(&done.completion);
115 return done.executed ? done.ret : -ENOENT;
118 /* This controls the threads on each CPU. */
119 enum multi_stop_state {
120 /* Dummy starting state for thread. */
122 /* Awaiting everyone to be scheduled. */
124 /* Disable interrupts. */
125 MULTI_STOP_DISABLE_IRQ,
126 /* Run the function */
132 struct multi_stop_data {
135 /* Like num_online_cpus(), but hotplug cpu uses us, so we need this. */
136 unsigned int num_threads;
137 const struct cpumask *active_cpus;
139 enum multi_stop_state state;
143 static void set_state(struct multi_stop_data *msdata,
144 enum multi_stop_state newstate)
146 /* Reset ack counter. */
147 atomic_set(&msdata->thread_ack, msdata->num_threads);
149 msdata->state = newstate;
152 /* Last one to ack a state moves to the next state. */
153 static void ack_state(struct multi_stop_data *msdata)
155 if (atomic_dec_and_test(&msdata->thread_ack))
156 set_state(msdata, msdata->state + 1);
159 /* This is the cpu_stop function which stops the CPU. */
160 static int multi_cpu_stop(void *data)
162 struct multi_stop_data *msdata = data;
163 enum multi_stop_state curstate = MULTI_STOP_NONE;
164 int cpu = smp_processor_id(), err = 0;
169 * When called from stop_machine_from_inactive_cpu(), irq might
170 * already be disabled. Save the state and restore it on exit.
172 local_save_flags(flags);
174 if (!msdata->active_cpus)
175 is_active = cpu == cpumask_first(cpu_online_mask);
177 is_active = cpumask_test_cpu(cpu, msdata->active_cpus);
179 /* Simple state machine */
181 /* Chill out and ensure we re-read multi_stop_state. */
183 if (msdata->state != curstate) {
184 curstate = msdata->state;
186 case MULTI_STOP_DISABLE_IRQ:
192 err = msdata->fn(msdata->data);
199 } while (curstate != MULTI_STOP_EXIT);
201 local_irq_restore(flags);
205 struct irq_cpu_stop_queue_work_info {
208 struct cpu_stop_work *work1;
209 struct cpu_stop_work *work2;
213 * This function is always run with irqs and preemption disabled.
214 * This guarantees that both work1 and work2 get queued, before
215 * our local migrate thread gets the chance to preempt us.
217 static void irq_cpu_stop_queue_work(void *arg)
219 struct irq_cpu_stop_queue_work_info *info = arg;
220 cpu_stop_queue_work(info->cpu1, info->work1);
221 cpu_stop_queue_work(info->cpu2, info->work2);
225 * stop_two_cpus - stops two cpus
226 * @cpu1: the cpu to stop
227 * @cpu2: the other cpu to stop
228 * @fn: function to execute
229 * @arg: argument to @fn
231 * Stops both the current and specified CPU and runs @fn on one of them.
233 * returns when both are completed.
235 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg)
237 struct cpu_stop_done done;
238 struct cpu_stop_work work1, work2;
239 struct irq_cpu_stop_queue_work_info call_args;
240 struct multi_stop_data msdata;
243 msdata = (struct multi_stop_data){
247 .active_cpus = cpumask_of(cpu1),
250 work1 = work2 = (struct cpu_stop_work){
251 .fn = multi_cpu_stop,
256 call_args = (struct irq_cpu_stop_queue_work_info){
263 cpu_stop_init_done(&done, 2);
264 set_state(&msdata, MULTI_STOP_PREPARE);
267 * If we observe both CPUs active we know _cpu_down() cannot yet have
268 * queued its stop_machine works and therefore ours will get executed
269 * first. Or its not either one of our CPUs that's getting unplugged,
270 * in which case we don't care.
272 * This relies on the stopper workqueues to be FIFO.
274 if (!cpu_active(cpu1) || !cpu_active(cpu2)) {
280 * Queuing needs to be done by the lowest numbered CPU, to ensure
281 * that works are always queued in the same order on every CPU.
282 * This prevents deadlocks.
284 smp_call_function_single(min(cpu1, cpu2),
285 &irq_cpu_stop_queue_work,
289 wait_for_completion(&done.completion);
291 return done.executed ? done.ret : -ENOENT;
295 * stop_one_cpu_nowait - stop a cpu but don't wait for completion
297 * @fn: function to execute
298 * @arg: argument to @fn
300 * Similar to stop_one_cpu() but doesn't wait for completion. The
301 * caller is responsible for ensuring @work_buf is currently unused
302 * and will remain untouched until stopper starts executing @fn.
307 void stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
308 struct cpu_stop_work *work_buf)
310 *work_buf = (struct cpu_stop_work){ .fn = fn, .arg = arg, };
311 cpu_stop_queue_work(cpu, work_buf);
314 /* static data for stop_cpus */
315 static DEFINE_MUTEX(stop_cpus_mutex);
316 static DEFINE_PER_CPU(struct cpu_stop_work, stop_cpus_work);
318 static void queue_stop_cpus_work(const struct cpumask *cpumask,
319 cpu_stop_fn_t fn, void *arg,
320 struct cpu_stop_done *done)
322 struct cpu_stop_work *work;
325 /* initialize works and done */
326 for_each_cpu(cpu, cpumask) {
327 work = &per_cpu(stop_cpus_work, cpu);
334 * Disable preemption while queueing to avoid getting
335 * preempted by a stopper which might wait for other stoppers
336 * to enter @fn which can lead to deadlock.
339 for_each_cpu(cpu, cpumask)
340 cpu_stop_queue_work(cpu, &per_cpu(stop_cpus_work, cpu));
344 static int __stop_cpus(const struct cpumask *cpumask,
345 cpu_stop_fn_t fn, void *arg)
347 struct cpu_stop_done done;
349 cpu_stop_init_done(&done, cpumask_weight(cpumask));
350 queue_stop_cpus_work(cpumask, fn, arg, &done);
351 wait_for_completion(&done.completion);
352 return done.executed ? done.ret : -ENOENT;
356 * stop_cpus - stop multiple cpus
357 * @cpumask: cpus to stop
358 * @fn: function to execute
359 * @arg: argument to @fn
361 * Execute @fn(@arg) on online cpus in @cpumask. On each target cpu,
362 * @fn is run in a process context with the highest priority
363 * preempting any task on the cpu and monopolizing it. This function
364 * returns after all executions are complete.
366 * This function doesn't guarantee the cpus in @cpumask stay online
367 * till @fn completes. If some cpus go down in the middle, execution
368 * on the cpu may happen partially or fully on different cpus. @fn
369 * should either be ready for that or the caller should ensure that
370 * the cpus stay online until this function completes.
372 * All stop_cpus() calls are serialized making it safe for @fn to wait
373 * for all cpus to start executing it.
379 * -ENOENT if @fn(@arg) was not executed at all because all cpus in
380 * @cpumask were offline; otherwise, 0 if all executions of @fn
381 * returned 0, any non zero return value if any returned non zero.
383 int stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
387 /* static works are used, process one request at a time */
388 mutex_lock(&stop_cpus_mutex);
389 ret = __stop_cpus(cpumask, fn, arg);
390 mutex_unlock(&stop_cpus_mutex);
395 * try_stop_cpus - try to stop multiple cpus
396 * @cpumask: cpus to stop
397 * @fn: function to execute
398 * @arg: argument to @fn
400 * Identical to stop_cpus() except that it fails with -EAGAIN if
401 * someone else is already using the facility.
407 * -EAGAIN if someone else is already stopping cpus, -ENOENT if
408 * @fn(@arg) was not executed at all because all cpus in @cpumask were
409 * offline; otherwise, 0 if all executions of @fn returned 0, any non
410 * zero return value if any returned non zero.
412 int try_stop_cpus(const struct cpumask *cpumask, cpu_stop_fn_t fn, void *arg)
416 /* static works are used, process one request at a time */
417 if (!mutex_trylock(&stop_cpus_mutex))
419 ret = __stop_cpus(cpumask, fn, arg);
420 mutex_unlock(&stop_cpus_mutex);
424 static int cpu_stop_should_run(unsigned int cpu)
426 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
430 spin_lock_irqsave(&stopper->lock, flags);
431 run = !list_empty(&stopper->works);
432 spin_unlock_irqrestore(&stopper->lock, flags);
436 static void cpu_stopper_thread(unsigned int cpu)
438 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
439 struct cpu_stop_work *work;
444 spin_lock_irq(&stopper->lock);
445 if (!list_empty(&stopper->works)) {
446 work = list_first_entry(&stopper->works,
447 struct cpu_stop_work, list);
448 list_del_init(&work->list);
450 spin_unlock_irq(&stopper->lock);
453 cpu_stop_fn_t fn = work->fn;
454 void *arg = work->arg;
455 struct cpu_stop_done *done = work->done;
456 char ksym_buf[KSYM_NAME_LEN] __maybe_unused;
458 /* cpu stop callbacks are not allowed to sleep */
465 /* restore preemption and check it's still balanced */
467 WARN_ONCE(preempt_count(),
468 "cpu_stop: %s(%p) leaked preempt count\n",
469 kallsyms_lookup((unsigned long)fn, NULL, NULL, NULL,
472 cpu_stop_signal_done(done, true);
477 extern void sched_set_stop_task(int cpu, struct task_struct *stop);
479 static void cpu_stop_create(unsigned int cpu)
481 sched_set_stop_task(cpu, per_cpu(cpu_stopper_task, cpu));
484 static void cpu_stop_park(unsigned int cpu)
486 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
487 struct cpu_stop_work *work;
490 /* drain remaining works */
491 spin_lock_irqsave(&stopper->lock, flags);
492 list_for_each_entry(work, &stopper->works, list)
493 cpu_stop_signal_done(work->done, false);
494 stopper->enabled = false;
495 spin_unlock_irqrestore(&stopper->lock, flags);
498 static void cpu_stop_unpark(unsigned int cpu)
500 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
502 spin_lock_irq(&stopper->lock);
503 stopper->enabled = true;
504 spin_unlock_irq(&stopper->lock);
507 static struct smp_hotplug_thread cpu_stop_threads = {
508 .store = &cpu_stopper_task,
509 .thread_should_run = cpu_stop_should_run,
510 .thread_fn = cpu_stopper_thread,
511 .thread_comm = "migration/%u",
512 .create = cpu_stop_create,
513 .setup = cpu_stop_unpark,
514 .park = cpu_stop_park,
515 .pre_unpark = cpu_stop_unpark,
519 static int __init cpu_stop_init(void)
523 for_each_possible_cpu(cpu) {
524 struct cpu_stopper *stopper = &per_cpu(cpu_stopper, cpu);
526 spin_lock_init(&stopper->lock);
527 INIT_LIST_HEAD(&stopper->works);
530 BUG_ON(smpboot_register_percpu_thread(&cpu_stop_threads));
531 stop_machine_initialized = true;
534 early_initcall(cpu_stop_init);
536 #ifdef CONFIG_STOP_MACHINE
538 int __stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
540 struct multi_stop_data msdata = {
543 .num_threads = num_online_cpus(),
547 if (!stop_machine_initialized) {
549 * Handle the case where stop_machine() is called
550 * early in boot before stop_machine() has been
556 WARN_ON_ONCE(msdata.num_threads != 1);
558 local_irq_save(flags);
561 local_irq_restore(flags);
566 /* Set the initial state and stop all online cpus. */
567 set_state(&msdata, MULTI_STOP_PREPARE);
568 return stop_cpus(cpu_online_mask, multi_cpu_stop, &msdata);
571 int stop_machine(int (*fn)(void *), void *data, const struct cpumask *cpus)
575 /* No CPUs can come up or down during this. */
577 ret = __stop_machine(fn, data, cpus);
581 EXPORT_SYMBOL_GPL(stop_machine);
584 * stop_machine_from_inactive_cpu - stop_machine() from inactive CPU
585 * @fn: the function to run
586 * @data: the data ptr for the @fn()
587 * @cpus: the cpus to run the @fn() on (NULL = any online cpu)
589 * This is identical to stop_machine() but can be called from a CPU which
590 * is not active. The local CPU is in the process of hotplug (so no other
591 * CPU hotplug can start) and not marked active and doesn't have enough
594 * This function provides stop_machine() functionality for such state by
595 * using busy-wait for synchronization and executing @fn directly for local
599 * Local CPU is inactive. Temporarily stops all active CPUs.
602 * 0 if all executions of @fn returned 0, any non zero return value if any
605 int stop_machine_from_inactive_cpu(int (*fn)(void *), void *data,
606 const struct cpumask *cpus)
608 struct multi_stop_data msdata = { .fn = fn, .data = data,
609 .active_cpus = cpus };
610 struct cpu_stop_done done;
613 /* Local CPU must be inactive and CPU hotplug in progress. */
614 BUG_ON(cpu_active(raw_smp_processor_id()));
615 msdata.num_threads = num_active_cpus() + 1; /* +1 for local */
617 /* No proper task established and can't sleep - busy wait for lock. */
618 while (!mutex_trylock(&stop_cpus_mutex))
621 /* Schedule work on other CPUs and execute directly for local CPU */
622 set_state(&msdata, MULTI_STOP_PREPARE);
623 cpu_stop_init_done(&done, num_active_cpus());
624 queue_stop_cpus_work(cpu_active_mask, multi_cpu_stop, &msdata,
626 ret = multi_cpu_stop(&msdata);
628 /* Busy wait for completion. */
629 while (!completion_done(&done.completion))
632 mutex_unlock(&stop_cpus_mutex);
633 return ret ?: done.ret;
636 #endif /* CONFIG_STOP_MACHINE */